Computer vision
Feature Matching Benchmark
A research benchmark for visual correspondence under controlled and natural variation, with component-level evidence and reproducible runs.
The project
Feature Matching Benchmark isolates detection, description, matching, and geometric verification to study where correspondence systems lose robustness. It evaluates classical and learned sparse feature pipelines on synthetic transformations and pinned HPatches sequences, preserving immutable observations and run provenance. The result is an operating-point comparison, not a universal leaderboard.
01
Component-level evidence
Separates feature extraction, matching, and homography verification so each result can be traced to a fixed pipeline configuration.
02
Held-out comparison
Compares ALIKED and XFeat on 120 HPatches pairs, measuring precision at five pixels alongside median CPU latency.
03
Reproducible analysis
Stores immutable Parquet observations with configuration, hashes, counts, and pinned pretrained weights for replayable analysis.
Built with
Primary
- Python
- NumPy
- OpenCV
Other
- ALIKED
- XFeat
- RANSAC
- LightGlue
- HPatches
- Parquet
- pytest